Publishing 2019 R1 content
[platform/upstream/dldt.git] / model-optimizer / extensions / middle / L2NormToNorm.py
1 """
2  Copyright (c) 2019 Intel Corporation
3
4  Licensed under the Apache License, Version 2.0 (the "License");
5  you may not use this file except in compliance with the License.
6  You may obtain a copy of the License at
7
8       http://www.apache.org/licenses/LICENSE-2.0
9
10  Unless required by applicable law or agreed to in writing, software
11  distributed under the License is distributed on an "AS IS" BASIS,
12  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  See the License for the specific language governing permissions and
14  limitations under the License.
15 """
16
17 import numpy as np
18
19 from mo.front.extractor import add_attrs_props
20 from mo.front.extractor import update_ie_fields
21 from mo.graph.graph import Node, Graph
22 from mo.middle.replacement import MiddleReplacementPattern
23
24
25 class L2NormToNorm(MiddleReplacementPattern):
26     enabled = True
27     force_clean_up = True
28
29     def run_after(self):
30         from extensions.middle.pass_separator import PreMiddleStart
31         return [PreMiddleStart]
32
33     def run_before(self):
34         from extensions.middle.pass_separator import MiddleStart
35         return [MiddleStart]
36
37     def pattern(self):
38         return dict(
39             nodes=[
40                 ('input', dict(kind='data')),
41                 ('l2_normalize', dict(kind='op', op='Mul')),
42                 ('l2_normalize_data', dict(kind='data')),
43                 ('maximum', dict(kind='op', op='Maximum')),
44                 ('maximum_data', dict(kind='data')),
45                 ('maximum_y_data', dict(kind='data')),
46                 ('rsqrt', dict(kind='op', op='Rsqrt')),
47                 ('rsqrt_data', dict(kind='data')),
48                 ('square', dict(kind='op', op='Square')),
49                 ('square_data', dict(kind='data')),
50                 ('sum', dict(kind='op', op='Reduce', reduce_type='sum')),
51                 ('sum_data', dict(kind='data')),
52             ],
53             edges=[
54                 ('input', 'square'),
55                 ('square', 'square_data'),
56                 ('square_data', 'sum'),
57                 ('sum', 'sum_data'),
58                 ('maximum_y_data', 'maximum'),
59                 ('sum_data', 'maximum'),
60                 ('maximum', 'maximum_data'),
61                 ('maximum_data', 'rsqrt'),
62                 ('rsqrt', 'rsqrt_data'),
63                 ('rsqrt_data', 'l2_normalize'),
64                 ('input', 'l2_normalize'),
65                 ('l2_normalize', 'l2_normalize_data'),
66             ]
67         )
68
69     def replace_pattern(self, graph: Graph, match: dict):
70         input_data_name = match['input'].node
71         output_data_name = match['l2_normalize_data'].node
72
73         if not match['maximum_y_data'].has_valid('value'):
74             return
75         if match['maximum_y_data'].value.shape != ():
76             return
77         y = match['maximum_y_data'].value
78
79         normalize_id = graph.unique_id()
80         graph.add_node(normalize_id,
81                        **add_attrs_props(
82                            dict(kind='op', precision="FP32", type='Normalize', name=str(graph.unique_id('normalize')),
83                                 op='Normalize', shape=None, eps=str(y), across_spatial=str(0), channel_shared=str(0),
84                                 data_type=None, infer=None, in_ports_count=2, out_ports_count=1)))
85         normalize_data_id = graph.unique_id()
86
87         graph.add_node(normalize_data_id, **add_attrs_props(graph.node[output_data_name]))
88         update_ie_fields(graph.node[normalize_id])
89         weights_id = graph.unique_id('weights_')
90         graph.add_node(weights_id, **add_attrs_props(
91             dict(kind='data', precision="FP32", name=weights_id, value=None, shape=None, data_type=None, infer=None)))
92         wnode = Node(graph, weights_id)
93         wnode['value'] = np.ones(shape=match['input'].shape[-1],
94                                  dtype=match['input'].data_type)  # TODO feature dim instead of -1
95         wnode['shape'] = np.array(wnode['value'].shape)
96         output_edges = list(graph.out_edges(output_data_name, data=True))
97         graph.remove_edges_from([
98             (input_data_name, match['l2_normalize'].id),
99             (input_data_name, match['square'].id)
100         ])
101         graph.remove_edges_from(list(graph.out_edges(output_data_name)))
102         graph.remove_node(output_data_name)
103         graph.add_edge(input_data_name, normalize_id, **{'in': 0})
104         graph.add_edge(weights_id, normalize_id, **{'in': 1, 'bin': 'weights'})
105         graph.add_edge(normalize_id, normalize_data_id, **{'out': 0})
106         for data, owner, attr in output_edges:
107             graph.add_edge(normalize_data_id, owner, **attr)